What does Together AI do?

Together AI is an AI-native cloud platform that lets enterprises train and run open-source AI models at scale without managing GPU infrastructure. It hosts more than 200 open-source models, including DeepSeek, Nemotron, MiniMax, and Kimi, and provides serverless inference, dedicated clusters, batch inference, and fine-tuning APIs. Its proprietary inference engine delivers speeds it claims are 31% higher than the next-fastest open-source alternative for production coding agent workloads.

How much cheaper is open-source inference compared to closed models?

Together AI reports that its customers achieve between 6x and 60x lower inference costs compared to closed-model providers for equivalent performance. One named example is Decagon, a customer service AI company, which cut its inference bill sixfold after switching to Together AI's platform. The range reflects the variability of workloads: commodity tasks can be run at dramatically lower cost, while the most demanding frontier-level use cases see smaller but still significant savings.

Who led the Together AI Series C and why does the investor list matter?

Aramco Ventures led the $800M round, with NVIDIA, Vista Equity Partners, General Catalyst, Emergence Capital, Schneider Electric, Salesforce Ventures, and others participating. Aramco Ventures treats AI infrastructure like energy infrastructure, a signal that sovereign and industrial capital now views open-source inference as critical compute. NVIDIA's participation aligns its GPU revenue interest with the open-source layer rather than proprietary model vendors. Schneider Electric's investment explicitly links energy efficiency to AI inference cost, reflecting how compute and energy economics are converging.

What is speculative decoding and why does it matter for enterprise AI costs?

Speculative decoding is an inference acceleration technique where a smaller, faster draft model proposes tokens that a larger target model verifies in parallel. When the draft is well-aligned with the target, this delivers 1.5x to 3x throughput improvements without changing output quality. Together AI's ATLAS system extends this with adaptive speculative decoding that learns from live traffic patterns in real time, achieving up to 500 tokens per second on DeepSeek-V3.1, a 2.65x speedup over standard decoding. At enterprise scale, this directly reduces the cost per inference call, which compounds significantly across agentic workloads generating millions of tokens daily.

What should enterprise AI teams do differently after this funding round?

The immediate action is to run a cost comparison between your current closed-model API spend and an equivalent open-source configuration on a platform like Together AI. The second action is to audit which workloads in your agent stack actually require frontier-model capability versus which can be served by a well-tuned open-source model at a fraction of the cost. The third is to evaluate your lock-in risk: if your agentic applications depend entirely on a single proprietary provider, the Together AI round underscores that a credible, enterprise-grade alternative now exists at scale.

Together AI Raises $800M: Open-Source Inference Goes Enterprise

On July 1, 2026, Together AI announced an $800 million Series C at an $8.3 billion valuation, led by Aramco Ventures with participation from NVIDIA, Vista Equity Partners, General Catalyst, Emergence Capital, Schneider Electric, Salesforce Ventures, and others. The company’s annual bookings crossed $1.15 billion last quarter. Open-source AI model usage across the industry has tripled in twelve months.

For enterprise AI teams still routing all inference through a single closed provider, this round is a signal worth acting on.

What Together AI Is (and Is Not)

Together AI is not a foundation model lab. It has never built a proprietary large language model. Since its founding in 2022 by Vipul Ved Prakash, Ce Zhang, Percy Liang, and Tri Dao, the company has built the infrastructure layer for running other organizations’ open-source models cheaply and at scale.

The platform offers four service tiers: serverless inference (pay per token, no cluster management), dedicated inference (reserved capacity with stricter reliability guarantees), batch inference (up to 50% cost reduction for non-real-time workloads), and fine-tuning APIs for custom model training. Underneath all of it sits a custom inference engine, a proprietary speculative decoding system called ATLAS, and recent research contributions including FlashAttention-4 for NVIDIA Blackwell and Together Megakernel for kernel-level optimization.

The model catalog includes more than 200 open-weights models: DeepSeek, Nemotron, MiniMax, Kimi, and GLM, among others. The company claims its endpoints are 31% faster than the next-fastest open-source inference alternative for production coding agent workloads.

Its customers include Cursor, Cognition (the company behind Devin), Decagon, ElevenLabs, and Suno. It serves more than one million developers.

The Economic Case the Round Validates

The argument at the center of Together AI’s pitch is straightforward: closed frontier model pricing that looks manageable in a prototype becomes unsustainable in production.

When a company deploys AI agents that write code, resolve customer issues, analyze documents, or automate revenue workflows at scale, token consumption grows rapidly. Inference bills compound faster than product budgets. The response from the most cost-disciplined teams has been to replace closed APIs with open-source models wherever performance parity exists, and to use closed frontier models only where the quality gap genuinely justifies the price.

Together AI quantifies the savings its customers achieve at 6x to 60x lower cost versus comparable closed-model deployments, with equivalent or better performance on most tasks. Decagon, a customer service AI company, cut its inference spend sixfold. A McKinsey study cited in Together AI’s funding announcement found that nearly three-quarters of organizations expect to increase their use of open-source AI.

The economic case is not hypothetical. It is already showing up in bookings.

What the Investor Mix Signals

The composition of this round deserves more attention than the headline number.

Investor	Category	Signal
Aramco Ventures (lead)	Sovereign / energy capital	AI infrastructure as utility-grade critical resource
NVIDIA	GPU hardware vendor	Open-source inference drives long-term GPU demand more than proprietary model vendors
Schneider Electric	Industrial / energy infrastructure	AI compute cost and energy efficiency are converging
Vista Equity Partners	Enterprise software PE	Open-source inference is enterprise-grade at scale
General Catalyst	Growth VC	Infrastructure layer is the durable position in the AI stack

The participation of Aramco Ventures leading the round alongside Schneider Electric is the most telling signal. Both organizations have primary exposure to energy infrastructure. AI inference is energy-intensive. When sovereign energy capital and industrial efficiency capital converge on the same inference platform, they are making a bet that open-source AI compute will be managed like a utility, not a software subscription.

NVIDIA’s participation is a different kind of signal. Together AI runs almost entirely on NVIDIA hardware. NVIDIA investing in the open-source inference layer rather than a proprietary model lab suggests the company believes GPU demand will be driven by open-source workloads scaling across thousands of enterprises rather than by a handful of frontier labs. Infrastructure bets compound. This one is betting that open-source inference is where the tokens will flow.

The Platform Stack Matters More Than the Funding Number

The $800 million matters less than what it will build. Together AI has committed to scaling compute capacity roughly 50 times over the next five years, backed by secured commitments for more than 500 MW of compute capacity capitalized independently by investors.

But the near-term competitive differentiation is the inference stack, not raw capacity.

Together AI’s ATLAS system, an adaptive speculative decoding engine, adjusts its token-drafting behavior from live traffic in real time. Unlike static speculative decoding, which uses a fixed draft model trained offline, ATLAS learns as workload patterns shift. The result is up to 500 tokens per second on DeepSeek-V3.1 and 460 tokens per second on Kimi-K2, a 2.65x throughput improvement over standard decoding. At the cost structures involved in large-scale agent deployments, this throughput advantage translates directly to dollars.

The company has also shipped FlashAttention-4 for NVIDIA Blackwell and Together Megakernel, kernel-level optimizations that reduce compute overhead for high-volume workloads. For enterprise teams running agentic pipelines with multi-step reasoning, long document analysis, or extended agent traces, these are not abstract engineering wins. They affect the economics of every workflow.

What This Means for Enterprise AI Strategy

The Together AI round clarifies something that has been true operationally for months but is now validated at $8.3 billion in market value and $1.15 billion in annual bookings: the open-source inference layer is real, enterprise-grade infrastructure.

For enterprise teams building agentic systems, the practical implications are three.

First, the cost audit is overdue. If your organization routes all inference through a single closed provider and has not run a structured comparison against open-source alternatives for each major workload category, you are likely paying a significant premium for capability you may not need on most tasks. The adoption readiness gap in enterprise AI is often framed as a technology problem. It is just as often a cost structure problem.

Second, model diversity reduces strategic risk. Together AI’s platform offers more than 200 models. No single model wins every task, and the open-source field is moving faster than most enterprise procurement cycles. Teams locked into a single provider have no flexibility when a better model ships. A platform approach lets teams route workloads to the appropriate model for each job without renegotiating contracts.

Third, the agentic cost curve is the one that matters. A single API call to a frontier model for a user query is manageable. An agentic workflow that generates thousands of calls per customer per day is a different calculation entirely. The forward-deployed engineering model that AWS, Microsoft, OpenAI, and Anthropic are all pursuing attempts to solve the deployment complexity problem. Together AI is solving the inference economics problem. Both problems are real, and solving one without the other leaves money on the table.

“Intelligence is becoming a foundational resource for the modern economy, every bit as essential as electricity, bandwidth, or capital,” said Vipul Ved Prakash, co-founder and CEO of Together AI in the company’s funding announcement. “Our mission is to ensure that intelligence is abundant, not expensive.”

For enterprise AI teams planning their next 12 months of infrastructure decisions, that framing is worth taking seriously. The infrastructure layer that makes AI economically viable at agent scale has just raised the capital to operate at utility scale.

The next step for most enterprise teams is not philosophical. It is a cost comparison. If you are running agentic GTM or operational AI systems at meaningful scale and have not benchmarked open-source inference costs against your current API spend, the Together AI round is the clearest possible prompt to do it now.

Sources: Together AI Series C announcement (primary), BusinessWire press release, TechCrunch, TechFundingNews, Together AI ATLAS research.